Measurements are never perfectly exact, so any quantity calculated from measurements also has uncertainty. Calculus gives a fast way to estimate how small input errors affect an output. The key idea is sensitivity, which tells how strongly a function responds to a small change in its input.
This matters in physics, engineering, chemistry, and data analysis whenever measured values are used in formulas.
For a computed quantity y = f(x), the differential dy = f'(x) dx approximates the output error caused by a small input error dx. If the function changes steeply near the measured value, a small measurement error can create a large output error. Relative error compares the error size to the value itself, and percentage error expresses that comparison as a percent.
For functions of several variables, partial derivatives show how uncertainty from each input contributes to the final uncertainty.
Understanding Calculus: Sensitivity and Error Propagation
The differential estimate comes from the tangent line idea. Near one measured input, a smooth curve behaves almost like its tangent line. The slope of that line converts a small horizontal shift into an estimated vertical shift.
This is why the method works best when the input uncertainty is small and the function is smooth over the whole likely measurement range. It becomes less reliable near sharp corners, discontinuities, turning points, or places where the formula is undefined.
A derivative of zero needs care too. The first order estimate may predict almost no output change, while curvature can still create a noticeable change.
Consider finding the area of a circle from a measured radius. Area depends on the radius squared, so a small fractional error in radius becomes about twice as large as a fractional error in area. If the radius is measured as ten centimetres with an uncertainty of one tenth of a centimetre, the radius has about a one percent uncertainty.
The area then has about a two percent uncertainty. The exact area is not the important part here. The important pattern is that squaring magnifies relative uncertainty.
A cube would magnify it by about three times. This appears in physics when volume is calculated from a measured length, because volume often depends on a length cubed.
Several measurements can affect one result in different ways. In a lab, density is found from mass divided by volume. An uncertain mass changes the density in one direction, while an uncertain volume changes it in the opposite direction.
For a cautious worst case estimate, treat each contribution as if it pushes the result as far as possible and add their sizes. This gives a useful upper limit. In repeated experiments, errors are often random rather than all pushing the same way.
Scientists may then combine independent random uncertainties by adding their squares and taking the square root. That method usually gives a smaller value than the worst case method, but it requires a justified assumption that the errors are independent.
Sensitivity has a practical meaning beyond calculation. It helps decide which measurement deserves more care. If changing the temperature slightly makes a calculated result change a lot, a more accurate thermometer may matter more than a more accurate ruler.
It can prevent wasted effort on quantities that barely affect the answer. Engineers use this idea when setting manufacturing tolerances.
Chemists use it when preparing solutions. Programmers use related ideas to check whether a model is stable when data are rounded or noisy.
When solving error problems, keep units attached to absolute uncertainties. A length uncertainty might be in millimetres, while a relative uncertainty has no unit. Round the final uncertainty to one or two sensible significant figures, then round the reported value to the same decimal place.
State whether the estimate is a maximum possible error or a random uncertainty, since those mean different things. Finally, check whether the uncertainty is small compared with the measurement. If it is not small, the tangent line approximation can be poor, and calculating the output at the upper and lower input limits is often safer.
Key Facts
- For one variable, dy = f'(x) dx estimates the change in y caused by a small change dx.
- Absolute error in y is often estimated by |dy| = |f'(x)| |dx|.
- Relative error is approximately |dy| / |y|, where y = f(x).
- Percentage error is relative error times 100%, so percentage error = (|dy| / |y|) x 100%.
- For y = f(x1, x2, ..., xn), the maximum estimated absolute error is |dy| <= |fx1| |dx1| + |fx2| |dx2| + ... + |fxn| |dxn|.
- For products and powers, relative errors often combine simply, such as if y = x^n then |dy| / |y| ≈ |n| |dx| / |x|.
Vocabulary
- Differential
- A differential is a small change estimate, such as dy = f'(x) dx, used to approximate how a function changes.
- Absolute error
- Absolute error is the estimated size of the uncertainty in a value, measured in the same units as the value.
- Relative error
- Relative error is the absolute error divided by the magnitude of the value being measured or computed.
- Percentage error
- Percentage error is the relative error multiplied by 100%.
- Sensitivity
- Sensitivity describes how much the output of a function changes in response to a small change in an input.
Common Mistakes to Avoid
- Using dy as the exact error, which is wrong because dy is a linear approximation that is best for small errors near the measured value.
- Forgetting the absolute value in error estimates, which is wrong because error size should be nonnegative even if the derivative is negative.
- Confusing absolute error with relative error, which is wrong because absolute error has units while relative error is a unitless fraction.
- Applying percentage error before computing the output value, which is wrong because percentage error in y must compare the estimated output error to |y|.
Practice Questions
- 1 A cube has side length x = 5.00 cm with possible error dx = 0.02 cm. Using V = x^3 and differentials, estimate the absolute error and percentage error in the volume.
- 2 The period of a pendulum is modeled by T = 2π sqrt(L / g). If g is treated as exact, L = 0.800 m, and dL = 0.004 m, estimate the relative error and percentage error in T.
- 3 A function has a very small derivative near the measured input value. Explain how that affects the propagated error in the computed output, and give a physical or mathematical example.